Bayesian methods for online real-time 3D imaging in challenging environments using single-photon LiDAR data

Abstract

Single-photon LiDAR (SPL) continues to gain interest in a variety of different applications. With LiDAR technology being deployed more outside of lab based conditions, it is critical to investigate methods for providing real-time scene reconstruction while reducing, in a principled way, the effects of noise and uncertainties caused by photon scattering environments, which is the aim of this thesis. Traditional 3D ranging methods for SPL usually perform surface detection and range estimation sequentially, alleviating the computational burden of joint detection and estimation. Furthermore, traditional approaches construct and process detected photon time of arrival (ToA) histograms to obtain final target depth estimates. However processing large data volumes over long temporal sequences results in undesirable costs in memory requirement and computational time. Adopting a Bayesian formalism, the initial joint detection/estimation problem is formulated as a single inference problem. Intractable integrals involved with variable marginalization in the Bayesian calculations are avoided by discretising variables, recasting the resulting problem as a model selection/averaging problem. A further approach is then investigated by using online Assumed Density Filtering (ADF) strategies to process SPL data on-chip without the need for histogram data construction. Additional benefits of the proposed methods are demonstrated by providing a conservative approach to uncertainty quantification of the calculated depth estimates, and real time analysis from the results. Statistical approaches can be limited by user defined input parameters and prior information. Finally, an approach is proposed using recursive Bayesian estimation to implement a detect-and-track method to SPL data processing which incorporates the inference information obtained from the previously mentioned joint detection/estimation approach. To avoid intractable calculations when computing the model parameters, a spatio-temporal correlation approach is proposed between individual model parameters to improve the quality of scene reconstruction. The benefits of the proposed methods are illustrated using synthetic, real SPL data for outdoor targets at up to 8.6 km as well as real data of underwater targets at up to 7.5 attenuation lengths from the LiDAR system.